Taking Contingency Seriously in Sensory-based Models of Learning in Infants

We hypothesize that contingency detection mechanisms are innate in young infants (e.g., Bahrick et al., 2002; Gergely & Watson, 1999) and propose to create sensory-based formal models of contingency learning. While other projects have implicitly used contingency mechanisms (e.g., Metta & Fitzpatrick, 2003), a focus on this topic in formal modeling appears rare. By sensory-based we mean that the models use sensors (i.e., input devices) such as digital cameras and microphones. We are using a contingency detection algorithm that computes the redundancy between a visual channel and another information channel (Hershey & Movellan, 2000). Based on this algorithm, ongoing projects involve detecting visual changes produced by commands sent to an animated face and an animated ball, and segmenting faces in a video of someone talking. Extending this work, we plan to combine contingency detection with learning algorithms to model uni-modal visual segmentation and other aspects of infant contingency-based learning